EVOKE Elicits Pretrained World Knowledge for Transferable Decision-Making via Fixed-State Goal Diversity

核心背景与行业痛点 LLM agents struggle when transferring to unfamiliar environments. Conventional world-model approaches predict future visual/state observations, suffering from high training burdens and compounding roll-out errors. Because digital world knowledge is already internalized within base LLMs during pretraining, the core hurdle lies in post-training: single-goal supervision per state drives agents toward superficial contextual habits rather than causal reasoning. ### 架构亮点与底层机制 EVOKE supplies targeted pressure to elicit latent world knowledge through goal diversity at fixed states: 1. Fixed-State Counterfactual Objectives: Holds the environmental state and conversation history invariant while introducing alternative goals. 2. Action Preference Re-ranking: Compels the agent to re-rank identical candidate actions under varying goals. Superficial context-matching fails under this regime, forcing the policy to simulate underlying state-action transitions. 3. Direct Decision Supervision: Bypasses explicit observation prediction networks, activating latent world representations directly through preference alignment objectives. ### 权威 Benchmark 与实测跑分对比 Evaluated across multiple multi-step agent decision benchmarks across three model backbones: 1. 14.3% Boost on Unseen Environments: EVOKE lifts task success on completely unseen environments by 14.3 percentage points over standard single-goal post-training while reducing redundant exploratory steps by 26%. 2. Exceptional Sample Efficiency: Matches and exceeds standard full-dataset baselines using only 40% of the trajectory training data. 3. Shortcut Suppression: Probing experiments confirm enhanced state-goal decoupling in representation space, eliminating vulnerability to misleading prompt artifacts. ### 开发者实战落地与开箱指南 EVOKE is open-sourced on GitHub with modular loss implementations and data augmentation scripts. Practitioners building enterprise RPA agents or cross-platform assistants can integrate fixed-state counterfactual goal swapping into existing DPO pipelines to unlock robust out-of-distribution transfer without adding inference latency.